Haroon MukhtarSenior Full Stack Engineer · AI Engineer

Challenge

During earnings season, analysts monitor live events, cross-reference internal reports, draft daily digests, and track KPI anecdotes by hand, across tools and data sources that do not talk to each other.

Solution

An earnings-intelligence system for a financial research firm. A self-healing daily monitor ingests earnings and event data, drafts digest emails, and updates calendars unattended. Alongside it run a 5-pass anecdote-generation pipeline and a research chat agent with semantic recall. Dedup and state live in code, and LLM calls are scoped to judgment tasks only.

Impact

A multi-agent financial research platform (TypeScript, Mastra) that automates earnings monitoring, analyst report drafting, and research chat. Deterministic code orchestrates the LLM reasoning, with live MCP integrations (Aiera, Gmail, Drive, Calendar) and a custom React front end.

Core features

• Self-healing daily Earnings & Events Monitor pulling live data via MCP
• Live MCP integrations with Aiera, Gmail, Google Drive, and Google Calendar
• Unattended synthesis of analyst-grade earnings digests and email drafts
• 5-pass LLM pipeline writing analyst anecdotes grounded in verbatim KPIs
• Research chat agent with semantic memory recall and per-request tool routing
• Custom React front end for research workflows and real-time interaction

Engineering highlights

• Mastra agent framework orchestrating LLM reasoning with deterministic TypeScript code
• LLM calls scoped to bounded judgment tasks, structured in and structured out
• State, deduplication, and templating owned by code rather than the model
• Daily calendar updates and email drafting run with no human in the loop

Challenge

Restaurant businesses needed fast, well-built websites without hiring engineers, and the platform behind those sites had to handle publishing, collaboration, SEO, accessibility, and hosting across many clients at once.

Solution

A no-code builder with a drafting and templating engine, Live and Staging publishing workflows, real-time collaborative editing, and a publishing pipeline that provisions domains and SSL on AWS automatically. I led the migration from Gatsby to Next.js for a rendering architecture that scales better.

Impact

Website delivery became self-serve. A restaurant site goes from draft to a live, accessible, SEO-ready deployment without an engineer in the loop.

Core features

• Live / staging publishing workflows with safe promotion
• Collaborative multi-user editing environment
• Drafting & templating engine for rapid site creation
• Analytics dashboard on GA4 + Google Search Console APIs
• Media library with editing and an optimization pipeline
• Independent mobile-specific CSS override system
• Domain & SSL configuration automation on AWS

Engineering highlights

• Gatsby → Next.js migration for scalable architecture
• ADA-compliant rendering across generated sites
• Strong Core Web Vitals / Lighthouse focus on published output
• Publishing workflow designed to scale across many client sites

Challenge

A general chatbot can discuss markets but cannot touch a user's actual portfolio. I wanted an assistant that reasons over live holdings, retrieves the context that matters, and runs multi-step workflows through real tools.

Solution

A LangChain-based agent with ReAct-style tool orchestration behind a FastAPI backend and a Next.js streaming chat frontend. ChromaDB powers RAG retrieval, persistent memory keeps conversations context-aware, and the agent plans multi-step reasoning chains across portfolio, pricing, and news tools.

Impact

An agent that observes, reasons, retrieves, and acts on real user data, on the architecture I would use in production.

Core features

• Streaming chat UI built with Next.js
• Ollama LLM integration behind FastAPI
• Portfolio analysis & market insights
• Portfolio CRUD tools the agent invokes directly
• Coin price lookup & financial news retrieval tools
• Persistent memory with context-aware multi-step reasoning

Engineering highlights

• LangChain agent framework with ReAct-style tool orchestration
• ChromaDB vector store powering the RAG pipeline
• Reasoning, retrieval, and tool execution kept in separate layers
• The agent plans and executes across several tools per request

Challenge

Growing businesses juggle purchase orders, shipments, invoices, and financial records across disconnected tools, losing visibility and duplicating work.

Solution

One platform for those workflows. Domains like Inventory, Finance, and CRM live as separate modules inside a monorepo, so the system stays maintainable while the workflows themselves stay connected.

Impact

One system of record for operations, from purchase through payment, in place of scattered spreadsheets and manual handoffs.

Core features

• Purchase orders, shipping, and invoicing workflows
• Financial workflows & reporting
• Vendor & customer management
• Activity tracking & operational automation

Engineering highlights

• Modular monorepo architecture
• Separated domains: Inventory, Finance, CRM
• Structured to stay maintainable as operations grow

Challenge

Video calls on their own left remote classes without scheduling, interactive teaching tools, or workflows for the different roles in a school.

Solution

An LMS with real-time classroom infrastructure: class scheduling, video and screen sharing, collaborative whiteboarding with annotations, and live chat, plus a dedicated portal for each role in the organization.

Impact

Teachers ran structured remote lessons using tools a plain video call does not provide.

Core features

• Class scheduling & management
• Video conferencing & screen sharing
• Whiteboarding with live annotations
• Real-time chat
• Teacher, sales, and admin portals

Engineering highlights

• Real-time multi-user infrastructure
• Role-based product surfaces on a shared platform